Skip to contentSkip to conversation

bpf_trace: build error without PERF_EVENTS

PATCH In review Arnd Bergmann 4 participants 5 messages Nov 10, 2015

Subsystems (6)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 bpf_trace: build error without PERF_EVENTS

In my ARM randconfig tests, I'm getting a build error for newly added code in bpf_perf_event_read and bpf_perf_event_output whenever CONFIG_PERF_EVENTS is disabled:

kernel/trace/bpf_trace.c: In function 'bpf_perf_event_read':
kernel/trace/bpf_trace.c:203:11: error: 'struct perf_event' has no member named 'oncpu' if (event->oncpu != smp_processor_id() ||

         ^
kernel/trace/bpf_trace.c:204:11: error: 'struct perf_event' has no member named 'pmu'
      event->pmu->count)

This can happen when UPROBE_EVENT is enabled but KPROBE_EVENT is disabled. I'm not sure if that is a configuration we care about, otherwise we could prevent this case from occuring by adding Kconfig dependencies.

Simply hiding the broken code inside #ifdef CONFIG_PERF_EVENTS as this patch does seems to reliably fix the error as well, I have built thousands of randconfig kernels since I started seeing this and added the workaround.

Fixes 62544ce8e01c ("bpf: fix bpf_perf_event_read() helper")
Fixes a43eec304259 ("bpf: introduce bpf_perf_event_output() helper")

---
I suspect my patch is not the right answer, but could someone please fix this?

kernel/trace/bpf_trace.c +4 -0
index 4228fd3682c3..82e0bc9d002a 100644
@@ -186,6 +186,7 @@ const struct bpf_func_proto *bpf_get_trace_printk_proto(void)
 	return &bpf_trace_printk_proto;
 }
 
+#if IS_ENABLED(CONFIG_PERF_EVENTS)
 static u64 bpf_perf_event_read(u64 r1, u64 index, u64 r3, u64 r4, u64 r5)
 {
 	struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
@@ -263,6 +264,7 @@ static const struct bpf_func_proto bpf_perf_event_output_proto = {
 	.arg4_type	= ARG_PTR_TO_STACK,
 	.arg5_type	= ARG_CONST_STACK_SIZE,
 };
+#endif
 
 static const struct bpf_func_proto *kprobe_prog_func_proto(enum bpf_func_id func_id)
 {
@@ -289,10 +291,12 @@ static const struct bpf_func_proto *kprobe_prog_func_proto(enum bpf_func_id func
 		return bpf_get_trace_printk_proto();
 	case BPF_FUNC_get_smp_processor_id:
 		return &bpf_get_smp_processor_id_proto;
+#if IS_ENABLED(CONFIG_PERF_EVENTS)
 	case BPF_FUNC_perf_event_read:
 		return &bpf_perf_event_read_proto;
 	case BPF_FUNC_perf_event_output:
 		return &bpf_perf_event_output_proto;
+#endif
 	default:
 		return NULL;
 	}
maintainer
#1 lore raw
Quoted 7 lines · Arnd Bergmann

In my ARM randconfig tests, I'm getting a build error for newly added code in bpf_perf_event_read and bpf_perf_event_output whenever CONFIG_PERF_EVENTS is disabled:

kernel/trace/bpf_trace.c: In function 'bpf_perf_event_read':
kernel/trace/bpf_trace.c:203:11: error: 'struct perf_event' has no member named 'oncpu' if (event->oncpu != smp_processor_id() ||

Quoted 8 lines
          ^
kernel/trace/bpf_trace.c:204:11: error: 'struct perf_event' has no member named 'pmu'
       event->pmu->count)

This can happen when UPROBE_EVENT is enabled but KPROBE_EVENT is disabled. I'm not sure if that is a configuration we care about, otherwise we could prevent this case from occuring by adding Kconfig dependencies.

I think that seems better than spreading #if IS_ENABLEDs into the code. Probably enough to add a 'depends on PERF_EVENTS' to config BPF_EVENTS, so it's also explicitly documented.

Quoted 8 lines

Simply hiding the broken code inside #ifdef CONFIG_PERF_EVENTS as this patch does seems to reliably fix the error as well, I have built thousands of randconfig kernels since I started seeing this and added the workaround.

Fixes 62544ce8e01c ("bpf: fix bpf_perf_event_read() helper")
Fixes a43eec304259 ("bpf: introduce bpf_perf_event_output() helper")

Thanks,
Daniel

maintainer
#2 lore raw
Quoted 12 lines · On Tue, 10 Nov 2015 14:31:38 +0100 Daniel Borkmann <[email protected]> wrote:
Quoted 11 lines · Arnd Bergmann

In my ARM randconfig tests, I'm getting a build error for newly added code in bpf_perf_event_read and bpf_perf_event_output whenever CONFIG_PERF_EVENTS is disabled:

kernel/trace/bpf_trace.c: In function 'bpf_perf_event_read':
kernel/trace/bpf_trace.c:203:11: error: 'struct perf_event' has no member named 'oncpu' if (event->oncpu != smp_processor_id() ||

          ^
kernel/trace/bpf_trace.c:204:11: error: 'struct perf_event' has no member named 'pmu'
       event->pmu->count)
Quoted 9 lines
Quoted 4 lines

This can happen when UPROBE_EVENT is enabled but KPROBE_EVENT is disabled. I'm not sure if that is a configuration we care about, otherwise we could prevent this case from occuring by adding Kconfig dependencies.

I think that seems better than spreading #if IS_ENABLEDs into the code. Probably enough to add a 'depends on PERF_EVENTS' to config BPF_EVENTS, so it's also explicitly documented.

So just do the following then?

-- Steve

kernel/trace/Kconfig +1 -1
index 8d6363f42169..f5aecff2d243 100644
@@ -434,7 +434,7 @@ config UPROBE_EVENT
 
 config BPF_EVENTS
 	depends on BPF_SYSCALL
-	depends on KPROBE_EVENT || UPROBE_EVENT
+	depends on KPROBE_EVENT && UPROBE_EVENT
 	bool
 	default y
 	help
#3 lore raw
Quoted 34 lines · Steven Rostedt
Quoted 12 lines · On Tue, 10 Nov 2015 14:31:38 +0100 Daniel Borkmann <[email protected]> wrote:
Quoted 11 lines · Arnd Bergmann

In my ARM randconfig tests, I'm getting a build error for newly added code in bpf_perf_event_read and bpf_perf_event_output whenever CONFIG_PERF_EVENTS is disabled:

kernel/trace/bpf_trace.c: In function 'bpf_perf_event_read':
kernel/trace/bpf_trace.c:203:11: error: 'struct perf_event' has no member named 'oncpu' if (event->oncpu != smp_processor_id() ||

          ^
kernel/trace/bpf_trace.c:204:11: error: 'struct perf_event' has no member named 'pmu'
       event->pmu->count)
Quoted 9 lines
Quoted 4 lines

This can happen when UPROBE_EVENT is enabled but KPROBE_EVENT is disabled. I'm not sure if that is a configuration we care about, otherwise we could prevent this case from occuring by adding Kconfig dependencies.

I think that seems better than spreading #if IS_ENABLEDs into the code. Probably enough to add a 'depends on PERF_EVENTS' to config BPF_EVENTS, so it's also explicitly documented.

So just do the following then?

-- Steve

kernel/trace/Kconfig +0 -0
index 8d6363f42169..f5aecff2d243 100644

kernel/trace/Kconfig config UPROBE_EVENT · line 437

Quoted 5 lines
kernel/trace/Kconfig +1 -1
@@ -434,7 +434,7 @@ config UPROBE_EVENT
 
 config BPF_EVENTS
 	depends on BPF_SYSCALL
-	depends on KPROBE_EVENT || UPROBE_EVENT
+	depends on KPROBE_EVENT && UPROBE_EVENT

yeah that's definitely cleaner and avoids ifdef creep in the future.

maintainer
#4 lore raw
Quoted 36 lines · Alexei Starovoitov
Quoted 35 lines · Steven Rostedt
Quoted 12 lines · On Tue, 10 Nov 2015 14:31:38 +0100 Daniel Borkmann <[email protected]> wrote:
Quoted 11 lines · Arnd Bergmann

In my ARM randconfig tests, I'm getting a build error for newly added code in bpf_perf_event_read and bpf_perf_event_output whenever CONFIG_PERF_EVENTS is disabled:

kernel/trace/bpf_trace.c: In function 'bpf_perf_event_read':
kernel/trace/bpf_trace.c:203:11: error: 'struct perf_event' has no member named 'oncpu' if (event->oncpu != smp_processor_id() ||

           ^
kernel/trace/bpf_trace.c:204:11: error: 'struct perf_event' has no member named 'pmu'
        event->pmu->count)
Quoted 9 lines
Quoted 4 lines

This can happen when UPROBE_EVENT is enabled but KPROBE_EVENT is disabled. I'm not sure if that is a configuration we care about, otherwise we could prevent this case from occuring by adding Kconfig dependencies.

I think that seems better than spreading #if IS_ENABLEDs into the code. Probably enough to add a 'depends on PERF_EVENTS' to config BPF_EVENTS, so it's also explicitly documented.

So just do the following then?

-- Steve

kernel/trace/Kconfig +0 -0
index 8d6363f42169..f5aecff2d243 100644
@@ -434,7 +434,7 @@ config UPROBE_EVENT
 
Quoted 6 lines

kernel/trace/Kconfig config UPROBE_EVENT · line 437

Quoted 4 lines
kernel/trace/Kconfig +1 -1
  config BPF_EVENTS
  	depends on BPF_SYSCALL
-	depends on KPROBE_EVENT || UPROBE_EVENT
+	depends on KPROBE_EVENT && UPROBE_EVENT

yeah that's definitely cleaner and avoids ifdef creep in the future.

Agreed, that's better.